Deck Design Judge

Deck design judge

Scorecard process: tier → locate → metrics → (shots) → rubric → merge. Leading words: scorecard, gate, evidence, self-critique.

Steps

1. Pick the tier

Done when: tier is stated (default T1 if unsure).

2. Locate deck + brand tokens

Find the .html (or shots dir). Read nearby colors_and_type.css / token files / logo SVGs. Note typeface, accent, radii, flat vs elevated.

Swiirl / Untitled-UI lineage is flat — elevation shadows are a gate.

Workspace: <deck-dir>/.design-eval/<timestamp>/.

Done when: deck path and flat/elevated rule are recorded.

3. Metrics (every tier)

python3 <skill>/scripts/deck_metrics.py <deck.html> \
  [--tokens <colors_and_type.css>] -o <workspace>/metrics.json

Swiirl HTML (optional — swiirl-design is internal brand-token tooling; skip this step if you don't have it):

python3 <swiirl-design>/scripts/check_brand_html.py <deck.html> \
  -o <workspace>/brand_check.json

Interpret flags with metrics.md.

Done when: metrics.json exists and every gate id is listed.

4. Screenshots (T2 / T3 only)

bash <skill>/scripts/render_slides.sh <deck.html> <workspace>/shots/

Open each PNG — overflow, clipping, accent area, 3-second rule. No Chrome → fall back to T1 and note it.

Done when: every shot reviewed, or T1 fallback noted.

5. Rubric judge

Read rubric.md. Score all 10 dimensions 0–5 with evidence (slide #, metric, quoted headline). Prefer the lower anchor when merely “fine.” T3: run the real panel — scripts/judge_panel.py <deck.html> --metrics <workspace>/metrics.json --out <workspace>/judge.json (median per model; errored models excluded, not zeroed) — see judge-panel.md.

Emit <workspace>/judge.json.

Done when: every dimension has score + evidence.

6. Scorecard

python3 <skill>/scripts/scorecard.py <workspace>/metrics.json <workspace>/judge.json \
  -o <workspace>/scorecard.md --json <workspace>/scorecard.json --deck <deck.html>

Any gate → grade capped at C, label Not ready. Show scorecard.md in order: Verdict → Gates → Dimensions → Top fixes → Metrics appendix.

Share your scorecard: --json emits a portable scorecard.json (schema_version 1.0 — deck sha256, tier, panel models, per-model votes, gates, dimensions, grade, top fixes). It's the shareable artifact for the public benchmark; provenance and the wider ecosystem are in ATTRIBUTION.md.

Done when: user has the scorecard path.

Self-critique loop

After generating a deck (before handoff):

score → if gates clear AND overall ≥ TARGET → ship
      → else apply top 1–3 fixes (gates first), re-measure, re-score
      → ≤3 rounds; report before/after delta

Default TARGET = 85, zero gates. Re-run metrics (and shots in T2) every round — never assume a fix worked.

Done when: target met, rounds exhausted, or progress stalled — ship best + delta.

In-skill reference

• Gates = objective walls (text density, shadows in flat systems, contrast, craft, frameworks).
• Rubric = weighted subjective half; the number is a compass, the fixes are the point.
• Prefer the brand bundle’s fonts/accent/radii/flat rule over generic defaults when present.
• Anchors: calibration.md.

Scripts